Benchmark Radar
AI BENCHMARK PROFILE

PAWBench

General AIMultimodal Perception

Evaluates video generators as stochastic samplers of world dynamics across 50 scenarios, comparing the distribution of possible behaviors under identical initial observations and actions.

Released
2026-08-27
Readiness
Paper only
Primary field
General AI

Why it matters

Current evaluations assess single-video plausibility, missing whether models recover the correct distribution of physical behavior, a requirement for reliable world modeling.

Motivation

Recent video generation models are increasingly framed as world models.

Primary resources

Benchmark Radar records only publicly supported details and links back to primary sources for verification.